Understanding tropical deforestation, one pixel at a time

Designing effective policy responses to tropical deforestation requires an understanding of the reasons behind it, commonly called deforestation drivers. In the tropics, over 90% of deforestation is driven by agriculture [1], but the specific commodities differ substantially across regions. For example, palm oil has been the most prominent deforestation driver in Indonesia, while it is pasture in Brazil and small scale, shifting agriculture in Central Africa [1,2]. Drivers also change in time because of new policies, economic shifts or changes in agricultural practices. Therefore, to enable targeted measures, it is necessary to attribute each new deforestation event to a specific driver.

In a recent study published in ISPRS Journal of Photogrammetry and Remote Sensing, Pišl et al. [3] propose a new approach to map deforestation drivers across the tropics. They train a deep learning model to recognize eleven land use classes, including the seven forest-risk commodities: cattle, oil palm, soy, rubber, cocoa, coffee, and timber. The model uses time series of satellite images as primary input, complemented by two additional data modalities: geographic coordinates and national statistics on the production of forest-risk commodities. Deforestation drivers are strongly clustered in space, and therefore providing the model a notion of location and regional trends can improve its performance in cases where satellite imagery alone is not enough.

Figure caption: The deep learning model architecture proposed in the study

When evaluated on a held-out dataset sampled across the tropics, the model achieves 87% accuracy across all classes, with the additional modalities bringing a 10% improvement. An analysis of the model’s behaviour shows that it uses the additional modalities to recognize classes strongly associated with spatial patterns, such as palm oil or rubber. To recognize drivers that can be found anywhere in the tropics, such as mining, the model mostly relies on the satellite imagery. The dataset used to train the model, compiled from free and public sources, is made available online [4].

The study provides an automatic, scalable and repeatable method to attribute deforestation to specific drivers. Independent of national governments and other stakeholders, it can support a wide range of use cases with objective data. For example, it can be used to detect areas where forest-risk commodities are produced on recently deforested land, indicating non-compliance with regulations such as the EU Deforestation Regulation. It can also improve the quantification of carbon emissions from deforestation, which depend on the specific land use that follows deforestation [5]. Overall, it demonstrates how modern machine learning can support evidence-based actions to conserve and protect tropical forests.

[1] Pendrill, Florence, et al. “Disentangling the numbers behind agriculture-driven tropical deforestation.” Science 377.6611 (2022): eabm9267.

[2] Singh, Chandrakant, and U. Martin Persson. “Global patterns of commodity-driven deforestation and associated carbon emissions.” (2024).

[3] Pišl, Jan, et al. “Mapping land uses following tropical deforestation with location-aware deep learning.” ISPRS Journal of Photogrammetry and Remote Sensing 232 (2026): 578-593.

[4] Pišl, Jan, et al. “Post-deforestation Land Use in the Tropics (PLUTo).” Version 2, Zenodo, 2025, https://doi.org/10.5281/zenodo.17831353.

[5] Don, Axel, Jens Schumacher, and Annette Freibauer. “Impact of tropical land‐use change on soil organic carbon stocks–a meta‐analysis.” Global Change Biology 17.4 (2011): 1658-1670.